Machine Learning Additive Corrections in a Gray-Box Model for Air Temperature in the Southwest Amazon Rainforest: Dry and Wet Season Analysis

This study develops an integrative modeling approach combining gray-box modeling, machine learning (ML), and Principal Component Analysis (PCA) to understand and predict micrometeorological dynamics in the Amazon rainforest, specifically at the Jaru Biological Reserve in Rondônia, Brazil. Utilizing extensive data from dry and wet seasons, we initially employed PCA to identify key meteorological variables, highlighting temperature, net radiation, and relative humidity as dominant drivers of atmospheric variability. A physics-based "dry model," describing temperature dynamics primarily through radiative forcing and cooling processes, provided a baseline subsequently corrected by ML-derived additive terms. A fully connected neural network successfully learned nonlinear corrections, significantly reducing prediction errors by over 60% in the dry season and approximately 54% in the wet season. To enhance interpretability, neural corrections were approximated using polynomial regression, revealing dominant nonlinear temperature-radiation interactions in both dry and wet conditions, only varying the magnitude of each combination. Our findings demonstrate marked seasonal contrasts in radiative forcing, atmospheric clarity, and cooling efficiency, underscoring the need for seasonally tailored modeling strategies. This combined physics-informed and data-driven methodology offers reliable and interpretable models essential for managing environmental impacts amid increasing anthropogenic and climatic pressures on the Amazon.